Diagnostic Accuracy of Artificial Intelligence-Assisted Radiology in Detecting Pulmonary Nodules on Chest CT: A Narrative Review
DOI:
https://doi.org/10.64252/kqrfxm88Keywords:
Artificial intelligence; Deep learning; Pulmonary nodules; Chest CT; Lung cancer screening; Diagnostic accuracy.Abstract
Background- Lung cancer has consistently been the world leader in cancer mortality, and screening has significantly reduced mortality with early detection. Even so, routine radiologic interpretation of low-dose computed tomography (LDCT) scans was met with challenges of high false positives, reader fatigue, and difficulty in detecting non-palpable pulmonary nodules. Artificial intelligence (AI), or rather deep learning algorithms, has been proposed as a potential adjunctive strategy in enhancing diagnostic accuracy in the detection of pulmonary nodules.
Methods- A narrative review was initiated to assess the diagnostic accuracy of AI-based radiology systems to identify pulmonary nodules on chest CT scans. Ten recent peer-reviewed clinical papers published in the last decade were included. Demographic and methodological details, specifications of the AI system, and diagnostic accuracy metrics such as sensitivity, specificity, false-positive rates, and inter-observer variability were extracted. Quantitative findings and qualitative data were synthesized narratively.
Results- AI-aided radiological interpretation showed consistently better sensitivity than traditional radiologist-alone interpretation, with gains of 6% to over 30% between studies. Detection sensitivity for small, subtle, and subsolid nodules improved markedly, often to about 90–99% sensitivity with AI aid compared with about 52–65% without AI. Specificity varied more between studies, tending to fall modestly by about 4–6% as a result of better false-positive detection. Qualitative feedback showed less reading time, better inter-observer agreement, and improved diagnostic confidence in radiologists employing AI systems. Overall radiologist fatigue was decreased with AI integration and encouraged consistency of difficult case interpretation, especially in less experienced readers.
Conclusion- AI-aided radiology greatly enhanced sensitivity for pulmonary nodule detection on chest CT, particularly for subtle lesions, but with little variation in specificity. AI systems enabled greater diagnostic uniformity and shorter reading times. Clinical application in the future would involve sensitive and specific balancing to avoid false positives.
